Anthropic发布Claude Opus 5.5：任务成本降低40%

Anthropic Launches Claude Opus 5.5, Task Costs Drop by 40% AIbase基地 Published inAI News · 6 min read · Sep 23, 2026 7 Anthropic released Claude Opus 5.5, the first model in the Claude 5.5 series. Anthropic claims that the new model performs comparably to Claude Fable 5.1 on most tasks, but with a 40% lower typical cost and over a 30% increase in output speed under default settings. About two hours after the release, OpenAI also launched GPT-6 Sol and GPT-6 Luna, adding new members to the GPT-6 "universe." Cost Reduced by 40%, Speed Increased by 30%, Programming Tests Lead In the 9 tests officially released, Opus 5.5 outperformed Fable 5.1, Opus 5, GPT-6 Astra, and GPT-5.6 Sol in 7 of them. In three programming tests (Terminal-Bench 4.0, FrontierCode v1.1, CursorBench 4.0), it achieved the highest scores, which were 66.4%, 54.4%, and 57.8% respectively; however, it still lags behind GPT-6 Astra in business process and research agent tests. On the Intelligence ranking list by the third-party institution Artificial Analysis, Opus 5.5 ranked first with 58 points, ahead of Fable 5.1 and GPT-6 Astra with 53 points each. In terms of pricing, Opus 5.5 charges $4 (approximately 26.8 yuan) per million input tokens and $20 (approximately 134 yuan) per million output tokens, a 20% reduction from Opus 5, with a 60% decrease in cache read prices. Claude Code and Claude Platform also offer a fast mode, which is up to 2.5 times faster than the standard mode, with input costing $8 and output costing $40 per million tokens. Anthropic also increased the five-hour usage quota for multiple paid plans and offered subscribers an opportunity to reset their quota at a self-selected time. 9.5 Hours Rewriting HAProxy, Long Tasks Become the Selling Point The main focus of Opus 5.5's upgrade clearly leans towards long-duration tasks. Anthropic stated that an early tester completed a migration involving 680,000 lines of code in less than a day; reviewing and fixing a 200,000-line codebase took Opus 5.5 less than 3 hours, while Opus 5 required more than 20 hours and consumed 2.5 times as many tokens. In internal testing, both models were tasked with rewriting the load balancing software HAProxy from C to Rust, and the rewritten versions passed almost all regression tests—Opus 5.5 took 9.5 hours and had a cost 51% lower than Fable 5.1. The financial report test was even more challenging, requiring the model to retrieve and write quarterly performance reports from hard-to-find web copies, with automatic scoring checking numbers and citations one by one. As a result, 16 out of 18 reports generated by Opus 5.5 met the standards, while neither Fable 5.1 nor Opus 5 succeeded. In knowledge work, it scored 1846 on GDPval-AA v2.1, higher than Fable 5.1's 1735. In writing and communication, Opus 5.5 has been adjusted to prioritize presenting key information and provide more concise answers. In security assessments, its frequency of attempting to break through isolation boundaries is about 85% lower than Opus 5 or Claude Mythos 5.1, and its attack success rate in the Gray Swan prompt injection test was among the lowest, matching that of Fable 5.1. However, some professional tasks are still limited—most cybersecurity tasks are transferred to Opus 4.8 for processing. Opus 5.5 is now available on Claude and platforms such as Amazon Web Services, Google Cloud, and Microsoft Azure. Developers can call it using the identifier claude-opus-5-5 and continue to have the option of zero data retention. Claude Sonnet 5.5 and Haiku 5.5 are planned for release in the coming weeks. ClaudeOpus5.5AnthropicGPT-6SolProgrammingTest This article is from AIbase Daily Scan to view Welcome to the [AI Daily] column! This is your daily guide to exploring the world of artificial intelligence. Every day, we present you with hot topics in the AI field, focusing on developers, helping you understand technical trends, and learning about innovative AI product applications. —— Created by the AIbase Daily Team © Copyright AIbase Base 2024, Click to View Source -